Personalized physical examination period dynamic estimation system and method based on big data

Abstract

The invention discloses a personalized physical examination period dynamic estimation system and method based on big data, and relates to the technical field of medical health data processing and intelligent evaluation. The estimation method comprises the step of constructing an integrated implementation framework including a multi-source data integration link, a multi-party cooperative processing link, a three-step modeling evaluation link and a quantitative feedback adjustment link. According to the personalized physical examination cycle dynamic estimation system and method based on big data, a medical institution, wearable equipment, family genetic disease history and government affair medical treatment are integrated through a cross-domain data fusion platform, and a multi-dimensional evaluation basis covering physiology, behavior, heredity and resource collaboration is constructed; and then through a three-step modeling process of a hybrid model training engine, the problem of insufficient data is solved through similar group data adaptation, nonlinear correlation features are mined through multiple decision trees, a risk level is finally output, and a period is dynamically adjusted in combination with medical resource assessment.

Description

Technical Field

[0001] This invention relates to the field of medical and health data processing and intelligent assessment technology, and in particular to a personalized dynamic prediction system and method for physical examination cycles based on big data. Background Technology

[0002] With the increasing awareness of public health management, regular physical examinations have become a core means of disease prevention and health monitoring. Currently, the physical examination cycle setting schemes commonly used in clinical practice are mainly based on the general recommendations of authoritative medical guidelines for different age and gender groups. This model has certain advantages in terms of standardization and promotion, and can provide the public with a basic health monitoring reference framework.

[0003] However, with the increasing demand for refined health management, the limitations of this fixed-cycle model have become increasingly apparent. From a data support perspective, the evaluation basis of existing solutions is mostly limited to static health data stored by medical institutions, such as historical physical examination reports and diagnostic records. It is not difficult to find that this single data source is insufficient to cover the dynamic information of users' daily life behaviors, such as daily exercise intensity, sleep quality, and dietary structure. It also does not include government service data such as the difficulty of making appointments at physical examination institutions and the distribution of equipment resources. This means that the completeness of the evaluation dimensions is insufficient, which may lead to biases in the judgment of users' true health status.

[0004] From the perspective of data processing and collaboration, medical and health data is usually stored in different institutions such as hospitals, community health service centers, and wearable device manufacturers. Due to considerations of data privacy protection, data silos are common among these institutions. If cross-domain data sharing is to be achieved to improve the evaluation basis, traditional data transmission models are prone to the risk of user privacy leakage. Therefore, it can be inferred that existing technologies are difficult to achieve collaborative processing of data from multiple institutions while ensuring privacy and security, which further limits the comprehensiveness of the evaluation.

[0005] The adaptation to new user groups is also worth noting; new users often lack sufficient historical health data, while the training and operation of existing assessment models are highly dependent on individual historical data. This means that when facing new users, the model may experience a cold start due to insufficient data support, resulting in a significant decrease in the reliability of the prediction results.

[0006] From the perspective of the dynamic adaptability of the cycle, the existing scheme remains fixed for a long time once the physical examination cycle is determined. However, in reality, users' health status will continue to change with factors such as changes in lifestyle, implementation of intervention measures, and aging. The fixed cycle obviously cannot keep up with such dynamic changes in a timely manner. For users with rising health risks, the early detection of diseases may be delayed due to the cycle being too long. For low-risk users with stable health status, the unnecessary consumption of medical resources may be caused by the cycle being too short. To this end, we propose a personalized physical examination cycle dynamic prediction system and method based on big data. Summary of the Invention

[0007] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a personalized physical examination cycle dynamic prediction system and method based on big data, which solves five core problems in existing physical examination cycle prediction technologies: single data dimension, insufficient privacy protection, prominent cold start problem, fixed and rigid cycle, and lack of intervention and coordination.

[0008] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A method for dynamically predicting personalized health checkup cycles based on big data, the method comprising: An integrated implementation framework is constructed, consisting of multi-source data integration, multi-party collaborative processing, a three-step modeling process, and dynamic adjustment based on quantitative feedback. Each link is seamlessly connected through encrypted data transmission and parameter interaction. The cross-domain data fusion platform performs cross-domain data aggregation by combining time calibration and spatial matching, and encrypts and associates structured health records from medical institutions, real-time behavioral information from wearable devices, family health history information declared by users, and information on the accessibility of medical resources from government platforms. The aggregated data is then stored in an encrypted user health information database. A multi-data-holder collaborative processing architecture is implemented based on a multi-party collaborative processing center: each data holder completes data cleaning and feature extraction through local data processing nodes, and transmits processing results only through encrypted parameter interaction, thus resolving the contradiction between data dispersion and privacy protection. The hybrid model training engine performs a three-step modeling process: data adaptation for similar groups, multi-dimensional feature mining, and health risk level determination. It uses public health big data to solve the problem of insufficient new user data, and automatically triggers continuous model updates every 1,000 new user data entries or every 3 months to adapt to changes in user health status. The cycle optimization and intervention system combines health risk level, authoritative medical guidelines for basic physical examination cycles, and real-time medical resource availability to generate a user-specific initial physical examination cycle. A quantitative feedback dynamic adjustment mechanism is established through a cycle optimization and intervention system: targeted health intervention suggestions are pushed and the implementation effect is tracked. The health change information after the intervention is fed back to the hybrid model training engine to re-determine the health risk level, thereby dynamically adjusting the physical examination cycle, so that the cycle prediction accuracy is improved by more than 25% compared with traditional methods.

[0009] Preferably, the cross-domain data aggregation specifically includes: The cross-domain data fusion platform adopts internationally unified time calibration and geolocation tag matching to achieve data spatiotemporal unity: using internationally unified time as a benchmark, the real-time behavior information of wearable devices (including heart rate and walking steps) is corrected for time, with a correction deviation of no more than 1 second; by using administrative division codes and latitude and longitude coordinates, the spatial location information of medical resources is associated, with a matching error of no more than 50 meters; The cross-domain data fusion platform establishes a multi-source data credibility grading standard: the credibility of structured health records from medical institutions is 100%, serving as the core benchmark data; the credibility of real-time behavioral information from wearable devices is set at 85% to 95% according to the device certification level; and the credibility of user-reported family health history information, after cross-verification with medical institution data, is set at 70% to 80%. The cross-domain data fusion platform has established data conflict resolution rules: when there are contradictions between data from different sources, such as when the user's declared weight differs from the medical institution's measurement by more than 5%, the comprehensive value is calculated according to the credibility weight, and the conflicting data is marked and pushed to the manual verification port. The cross-domain data fusion platform uses a symmetric encryption algorithm to achieve low-power transmission of information from wearable devices and the WebSocket protocol to achieve high-frequency data real-time synchronization, ensuring that the data transmission delay does not exceed 3 seconds and that the encryption strength meets the national level 3 security requirements.

[0010] Preferably, the multi-data holder collaborative processing architecture includes: Data processing stage: Local data processing nodes of data holders such as hospitals, equipment manufacturers, and government platforms use medical knowledge databases for comparison and isolated data identification to filter and remove abnormal data. They also fill in missing data by using data from users of the same age and gender and matching historical data patterns to ensure that the local data integrity rate is not less than 98%. Feature processing stage: After the local data processing node extracts basic features, redundant features are removed by cross-domain feature principal component analysis, and the feature parameters are encrypted using an asymmetric encryption algorithm. The encrypted feature parameters are then uploaded to the multi-party collaborative processing center. Model training phase: The multi-party collaborative processing center distributes encrypted feature parameters to each sub-model training node: the hospital node trains the physiological indicator sub-model, the equipment manufacturer node trains the behavioral pattern sub-model, and the government node trains the resource adaptation sub-model; after each sub-model is trained, only the model parameters are uploaded to the multi-party collaborative processing center to be aggregated and integrated into a global evaluation model, without transmitting the original data throughout the process.

[0011] Preferably, the three-step modeling process includes: Similar group data adaptation stage: The hybrid model training engine selects a public health dataset with a demographic similarity of no less than 80% to the target user group, and adopts a sample weight dynamic adjustment method to adapt more than 70,000 labeled sample features to the target user group, ensuring that the model prediction accuracy is still no less than 75% when the number of target user samples is less than 100. Multi-dimensional feature mining stage: The hybrid model training engine uses 100 decision trees to mine features, with the decision tree depth set to 8 to 12 layers and the parameter adjustment step size set to 0.05 to 0.1. Non-linear correlation features such as blood pressure fluctuation, sleep duration and salt intake are mined, and the top 30 core features are selected by ranking the importance of features. In the health risk level determination stage: the hybrid model training engine inputs the core features into the risk assessment model and outputs a health risk value of 0 to 1, which is divided into high, medium and low risk levels according to the thresholds of 0.7 and 0.3. The model is validated by a five-fold cross-validation method, which requires that the model prediction matches the actual health status with a degree of not less than 0.88. Otherwise, the model parameters are re-optimized.

[0012] Preferably, the generation of the initial physical examination cycle includes: The system for optimizing and intervening in the cycle constructs a medical resource accessibility assessment value, which consists of appointment waiting time (weight 0.4), equipment utilization saturation (weight 0.3), and the nearest physical examination institution (weight 0.3). All three dimensions are converted to standardized values ​​from 0 to 1 and then summed with weights. The cycle optimization and intervention system implements the following basic physical examination cycle adjustment rules: For high-risk users with a risk value ≥ 0.7, the cycle = basic cycle × (0.5-0.7) × assessment value; for medium-risk users with a risk value ≤ 0.3, the cycle = basic cycle × 1.0 × assessment value; for low-risk users with a risk value ≥ 0.3, the cycle = basic cycle × (1.2-1.4) × assessment value. Example: For a 45-year-old male with a baseline cycle of 1 year, if the risk value is 0.8 and the assessment value is 0.9, the initial cycle = 1 year × 0.6 × 0.9 = 0.54 years, approximately 6.5 months. The generated results are synchronized to the user interaction terminal and the medical institution terminal.

[0013] Preferably, the quantitative feedback dynamic adjustment mechanism includes: The cycle optimization and intervention system calls the intervention suggestion library to generate solutions: for high-risk users with abnormal blood pressure, it generates suggestions such as walking briskly for 30 minutes a day, consuming ≤5 grams of salt per day, and measuring blood pressure twice a week, and pushes them to the user's mini-program; The cycle optimization and intervention system evaluates the intervention effect: constructs an intervention effect evaluation value, with the degree of improvement of indicators at 40%, the degree of completion of the plan at 30%, and subjective scores at 30%. The evaluation value is calculated by collecting execution data through wearable devices, subjective scores filled in by users, and follow-up data provided by medical institutions. The cycle optimization and intervention system dynamically adjusts the cycle: the cycle is extended by 10% to 20% for assessment values ​​≥ 8 points, the cycle remains unchanged for scores of 5 to 8 points, and the cycle is shortened by 15% to 30% for scores of 5 points, with the intervention plan optimized; every 3 months, the assessment results and updated data are sent back to the hybrid model training engine to re-determine the risk level and adjust the cycle, forming a cycle.

[0014] A personalized health check-up cycle dynamic prediction system based on big data, the system includes: Cross-domain data fusion platform: integrates multi-source data interface adaptation module, spatiotemporal alignment processing module, data credibility assessment module, conflict resolution module, and encrypted transmission module; it connects to medical institution systems through HL7FHIR interface, wearable devices through MQTT interface, and government affairs platform through RESTful interface; it uses ID card number plus fingerprint feature encryption value as the unique health identifier to build an encrypted user health information database, realizing the aggregation and storage of multi-source data; Multi-party collaborative processing center: Deployed on a cloud server, it includes a collaborative parameter scheduling module, an encrypted calculation module, and a model parameter aggregation module; it receives encrypted feature parameters from the local nodes of each data holder, schedules sub-model training, aggregates and generates a global evaluation model, and manages data interaction security throughout the process; Hybrid model training engine: It has built-in modules for similar group data adaptation, multi-decision tree feature mining, health risk level determination, model performance monitoring, and bias correction; it loads pre-trained models from public health datasets, performs incremental training after receiving new data, and triggers model updates every 3 months or with 1000 new samples to ensure prediction accuracy ≥85%; The cycle optimization and intervention system consists of an initial cycle generation module, an intervention suggestion push module, an effect evaluation and analysis module, an adjustment and execution module, and a user interaction module. It generates an initial cycle based on the output of a global evaluation model, pushes intervention suggestions and collects feedback data, and triggers cycle adjustment and model iteration. Security and compliance assurance system: integrates a sensitive information encryption and replacement module, a role-based access control module, and an operation behavior audit log module; uses the AES256 encryption algorithm to protect data, implements hierarchical authorization from users and doctors to administrators, and records data operations throughout the process.

[0015] Preferably, the spatiotemporal alignment processing module of the cross-domain data fusion platform includes: Time calibration unit: Receives time stamps from various data sources, converts them to internationally unified time, and corrects data with a deviation >10 seconds through linear interpolation; Spatial mapping unit: Converts medical institution and user addresses into latitude and longitude coordinates, matches them with the medical resource spatial database, and feeds the matching results back to the data credibility assessment module; Data storage unit: Redis cluster is used to store the real-time behavior information of wearable devices for the past 7 days, and HBase is used to store the full historical health information. Millisecond-level queries are supported, and data retention complies with the medical data retention period regulations, which is ≥5 years.

[0016] Preferably, the bias correction module of the hybrid model training engine includes: Deviation detection unit: Periodically calculates the KL divergence between the model's predicted results and the actual health outcomes. When the divergence is greater than 0.2, it is determined to be a model deviation and an early warning is triggered. Adaptive adjustment unit: If the data drift is due to a change in feature distribution, call the similar group data adaptation module to supplement the samples; if the concept drift is due to a change in risk definition, update the threshold and evaluation criteria of the health risk level judgment module. Model version management unit: Saves historical model parameters and supports version rollback; before a new model goes live, it undergoes five-fold cross-validation, and after successful validation, it replaces the old model to ensure iteration safety.

[0017] Preferably, the adjustment execution module of the cycle optimization and intervention system includes: Rule Management Unit: Employs the Drools rule engine to solidify the rules governing the correspondence between risk level, periodic adjustment coefficient, and intervention plan, and supports online updates; Data Feedback Unit: Receives intervention execution data from wearable devices, subjective health scores of 1 to 10 points submitted by users, and follow-up indicator data from medical institutions via API interface, and transmits them to the effect evaluation and analysis module; Visualization unit: Displays health risk values, current physical examination cycle, and intervention effect trend curves to users; pushes high-risk user warnings to medical institutions; and provides health departments with regional risk distribution and intervention effect statistical reports to support public health decision-making.

[0018] (III) Beneficial Effects 1. By integrating static health records from medical institutions, dynamic behavioral data from wearable devices, family genetic disease history, and government medical resource data through a cross-domain data fusion platform, a multi-dimensional assessment foundation covering physiology, behavior, genetics, and resource collaboration is constructed. Then, through a three-step modeling process using a hybrid model training engine, the problem of insufficient data is first addressed by adapting similar group data, and then nonlinear correlation features are mined through multi-decision trees. Finally, an accurate risk level is output, and the cycle is dynamically adjusted in conjunction with the medical resource accessibility assessment. It is precisely because of the completeness of multi-source data and the refinement of modeling that the cycle prediction accuracy of this invention is more than 25% higher than that of traditional fixed cycle methods. This not only avoids delays in disease detection for high-risk users due to excessively long cycles, but also reduces unnecessary physical examination costs for low-risk users, achieving precise health management tailored to each individual.

[0019] 2. Each data holder completes data cleaning and feature extraction at its local node, transmitting only the encrypted feature parameters to the center via an asymmetric encryption algorithm. The global evaluation model is generated by aggregating the parameters of each sub-model, without involving cross-domain transmission of the original health data. Simultaneously, a security and compliance assurance system forms a full-link privacy protection barrier through sensitive information encryption and replacement, role-based hierarchical authorization, and full recording of operation logs. The implementation of these technical measures effectively resolves the contradiction between data sharing and privacy protection, eliminating security risks associated with cross-domain data collaboration. Furthermore, by setting up an automatic update mechanism of adding 1000 new samples or every 3 months through the hybrid model training engine, it is ensured that the model can adapt to changes in user health data in real time.

[0020] 3. A complete health management system, encompassing assessment, intervention, feedback, and adjustment, is constructed through a cycle optimization and intervention system. This system pushes targeted intervention suggestions based on risk levels, evaluates intervention effectiveness using multi-source data, directly links with the model to reassess risks and adjust cycles. For cases with excellent pre-existing effects, the cycle is appropriately extended; for those with poor effects, the cycle is shortened and the intervention plan optimized. This deep synergy between cycle setting and health intervention transforms physical examinations from isolated procedures into a core node throughout the entire health management process, significantly improving the effectiveness of health interventions and user compliance. Furthermore, it connects seamlessly to medical institutions' HIS / LIS systems via the HL7FHIR standard interface, mainstream wearable devices via the MQTT protocol interface, and government platforms via a RESTful interface, achieving seamless integration without requiring reconstruction of existing medical, equipment, and government systems. The cycle optimization and intervention system displays risk levels, cycles, and intervention effects to users via a mini-program / app, pushes high-risk warnings to medical institutions, and provides regional health trend analysis to health departments, meeting the needs of multiple users. Attached Figure Description

[0021] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is an overall flowchart of an embodiment of the present invention. Detailed Implementation

[0023] This application provides a personalized health check-up cycle dynamic prediction system and method based on big data, addressing five core problems in existing health check-up cycle prediction technologies: single data dimension, insufficient privacy protection, prominent cold start problem, fixed and rigid cycles, and lack of intervention and coordination. It integrates static health records from medical institutions, dynamic behavioral data from wearable devices, family genetic history, and government medical resource data through a cross-domain data fusion platform, constructing a multi-dimensional assessment foundation covering physiology, behavior, genetics, and resource coordination. Then, through a three-step modeling process using a hybrid model training engine, it first adapts similar group data to address the data shortage problem, then mines nonlinear correlation features through multi-decision trees, and finally outputs an accurate risk level, dynamically adjusting the cycle based on medical resource accessibility assessment. Relying on the completeness of multi-source data and the refinement of modeling, the cycle prediction accuracy of this invention is more than 25% higher than traditional fixed-cycle methods. This avoids delays in disease detection for high-risk users due to excessively long cycles, and reduces unnecessary health check-up costs for low-risk users, achieving precise health management tailored to each individual.

[0024] Example: The technical solution in this application aims to address five core problems in existing physical examination cycle prediction technologies: limited data dimensions, insufficient privacy protection, prominent cold start issues, fixed and rigid cycles, and lack of intervention and coordination. The overall approach is as follows: With the advancement of the Healthy China strategy, regular physical examinations have become a core means of disease prevention. In current clinical practice, the examination cycle is mostly set according to the "Expert Consensus on Basic Items of Health Examination in China", such as once a year for adults aged 40 to 59 and once every six months for those over 60. This one-size-fits-all model based on age and gender has significant limitations: high-risk users, such as those with prehypertension, may experience delays in intervention due to the excessively long cycle, while low-risk users may waste medical resources due to the excessively short cycle.

[0025] While existing research attempts to optimize the health check cycle, it still suffers from four key shortcomings: First, the data dimensions are limited, relying heavily on static electronic health records (EHRs) from medical institutions and failing to integrate dynamic behavioral data from wearable devices, such as exercise, sleep, and family history of genetic diseases. Second, privacy protection is insufficient, as centralized data storage requires the transfer of raw data across institutions, which can easily lead to leakage risks. Third, it is difficult to adapt to new users, as the model relies on historical data, and the prediction accuracy drops sharply when new users are cold-started. Fourth, the cycle is rigid and lacks intervention coordination, adjusting only at fixed time intervals without dynamically optimizing in conjunction with the effects of health interventions.

[0026] To address the aforementioned issues, this paper proposes a personalized health check-up cycle dynamic prediction system and method based on big data. The core objectives are: to enhance the completeness of assessment dimensions through deep fusion of multi-source data; to ensure privacy and security through distributed modeling; to solve the cold start problem through adaptation to similar groups; and to achieve dynamic cycle optimization through intervention feedback loops. Ultimately, this forms a comprehensive technical solution covering the entire process from data, modeling, cycle prediction to intervention. Simultaneously, an engineering system architecture is designed to ensure the technology's practicality and applicability. To address the problems existing in the prior art, this invention provides a method for dynamically predicting personalized physical examination cycles based on big data. This prediction method includes: 1. Overall Technical Framework: Multi-source data integration serves as a prerequisite to address the issue of single data dimensions. Specifically, it involves using a cross-domain data fusion platform to collect data from four categories: HIS / LIS systems of medical institutions, smartwatches from wearable devices, user self-reports, and medical resource aggregation from government platforms. The data collection frequencies vary, with medical institutions updating in real time, wearable devices sampling at the minute level, and government data being updated at the hour level. The data format is unified as a combination of timestamp, data value, and source identifier to ensure compatibility in subsequent processing. Multi-party collaborative processing is used to solve privacy protection issues. Specifically, each data holder processes the data locally. Data holders include hospitals, equipment manufacturers, and government platforms. Only feature parameters are encrypted and transmitted to the collaboration center, which aggregates them to generate a global model. RSA 2048-bit asymmetric encryption is used. The collaboration center only aggregates the data after receiving all sub-model parameters to avoid modeling bias caused by missing data. The three-step modeling process is designed to address the issues of cold start and assessment accuracy. Specifically, it involves adapting to similar groups, mining multi-dimensional features, determining health risk levels, outputting risk values ​​from 0 to 1 and classifying them as high / medium / low risk, using the XGBoost algorithm to construct 100 decision trees, selecting 30 core features with a cumulative contribution of ≥90%, and ensuring an AUC value of ≥0.88 through five-fold cross-validation. Quantitative feedback and dynamic adjustment aim to address the issues of rigid cycles and intervention coordination. Specifically, it involves generating an initial cycle, pushing intervention suggestions, evaluating effectiveness, and finally adjusting the cycle, repeating every three months. The intervention program library contains over 100 standardized programs, updated every six months according to guidelines. Effectiveness evaluation uses a weighted calculation of indicator improvement (0.4), program completion (0.3), and subjective rating (0.3). Adjustment coefficients are categorized into three levels based on effectiveness: excellent (0.1-0.2), good (0), and poor (-0.15--0.3). As can be seen, by coordinating the three layers of objectives, operations, and implementation details, the integrated framework is transformed from an abstract concept into an executable technical process. Each step corresponds to detailed methods and system components in subsequent chapters, with no logical breakpoints.

[0027] 2. Process: (1) Cross-domain data aggregation: The process mainly consists of four steps: spatiotemporal unification, trust level classification, conflict resolution, and encrypted storage. The specific details are as follows: In terms of spatiotemporal consistency, time calibration uses the NTPv4 protocol to synchronize to UTC time. Wearable devices synchronize once per hour, and the hospital system synchronizes at 3:00 AM daily to avoid peak business hours. Spatial mapping uses the Gaode Map geocoding API to convert addresses into WGS84 latitude and longitude. For fuzzy addresses, such as XX Community, the location is supplemented by the user's frequently used medical examination institutions. In the example, user A's address is XX Road, Xihu District, Hangzhou City, Zhejiang Province, which is converted to 30.2741°N, 120.1699°E, and matches the nearby medical examination institution at 30.2792°N, 120.1751°E. The credibility grading adopts a three-level standard: Medical institution data has a credibility of 100%, such as blood pressure measured in hospitals; data from certified wearable devices has a credibility of 85% to 95%, such as exercise data from Huawei watches; and user-reported data has a credibility of 70% to 80%. For family medical history, cross-verification with hospital data is required. User-reported data without medical data verification, such as a grandfather with diabetes but no hospitalization record, has a credibility reduced to 60% and is marked as pending verification, to be upgraded after the user provides proof. Data conflict resolution uses the following formula: Where V is the comprehensive value of data conflicts, and the unit is the same as that of the original data; W i Let V be the credibility weight of the i-th data source, such as hospital data W1=1.0, user reports W2=0.75, and V iLet V be the original value of the i-th data source, such as V1=73kg measured by the hospital and V2=70kg reported by the user. N is the number of conflicting data sources, here N=2. For example, if user A reports a weight of 70kg, the corresponding weight W2 is 0.75, and the hospital measures 73kg, the corresponding weight W1 is 1.0. We can calculate V=(1.0×73+0.75×70) / (1.0+0.75)=71.7kg. We mark the conflict and push it to medical staff for review. Based on user A's recent exercise data, we can fine-tune it to 72kg.

[0028] Encrypted storage employs a tiered strategy, using hot, warm, and cold data tiers. Hot data, including real-time wearable device behavior information from the past 7 days, is stored in a Redis cluster; warm data, including structured health records from medical institutions from the past 3 months, is stored in a MySQL cluster; and cold data, including historical data from the past 3 months, is stored in HBase. Backups utilize master-slave synchronization, with daily incremental backups and weekly full backups. Backup data is stored in an off-site disaster recovery center. All sensitive information, such as ID numbers, is encrypted using the AES256 algorithm and backed up to the off-site disaster recovery center, which is at least 100 kilometers away from the main center, complying with the "Regulations on Medical Record Management in Medical Institutions." The backup strategy utilizes: ① Tools: NTP / Gaode API; ② Formula: Conflict calculation; ③ Example: User A address matching; ④ Compliance details: The backup strategy transforms cross-domain data aggregation from abstract steps into actionable technical operations, with each operation having specific parameters such as synchronization frequency, positioning accuracy, and storage latency, without ambiguity.

[0029] Multi-party collaborative processing: A distributed modeling architecture employs local processing and encrypted parameter aggregation to ensure the model moves while the data remains stationary. Local data processing serves as a preliminary step, with hardware configurations differentiated according to role. Hospital nodes utilize an 8-core CPU + 16GB RAM + 1TB SSD, supporting 100,000 data entries per day; equipment manufacturer nodes use a 4-core CPU + 8GB RAM + 512GB SSD, supporting 50,000 behavioral data entries per day; and government nodes use a 6-core CPU + 12GB RAM + 1TB HDD, supporting 10,000 resource data entries per day. A data cleaning toolkit is also deployed, including outlier detection, medical knowledge verification, and an isolated forest algorithm to remove outliers (e.g., blood pressure > 300 mmHg). Missing data is filled using the mean of similar users, combined with historical pattern matching (e.g., a 35-year-old woman's fasting blood glucose is filled with 5.2 mmol / L), ensuring data integrity ≥ 98%. Feature encryption and transmission are the core of privacy, employing RSA 2048-bit asymmetric encryption. Each data holder generates a unique public key to encrypt feature parameters, which, along with a private key, is used to decrypt aggregation parameters issued by the collaboration center. The parameter format includes a parameter identifier: physiological / behavioral / resource + encrypted string + timestamp. The timestamp is valid for 10 minutes to prevent replay attacks. Transmission uses the TLS 1.3 protocol, with port 443 to the collaboration center. Plaintext transmission of any sensitive information is prohibited. Real-time monitoring is implemented during transmission, and automatic retry is performed if the delay exceeds 10 seconds. Distributed model training and aggregation serve as the core of modeling. Sub-model training is divided by role: the physiological indicator sub-model for hospital nodes uses a logistic regression algorithm, with input features including blood pressure, blood sugar, and BMI, and outputting physiological risk weights; the behavioral pattern sub-model for equipment manufacturer nodes uses a random forest algorithm, with input features including exercise achievement rate and sleep disorder frequency, and outputting behavioral risk weights; the resource matching sub-model for government nodes uses a linear regression algorithm, with input features including appointment waiting time and equipment utilization rate, and outputting resource matching weights. Global aggregation uses formula P... global =α×P physio +β×P behavior +γ×P resource , where P global The core parameters of the global evaluation model are unitless and range from 0 to 1. α=0.5, β=0.3, and γ=0.2 are the weights of the physiological, behavioral, and resource sub-models, respectively, set based on clinical importance. P physio For physiological indicator sub-model parameters, such as user A's P physio =0.75, P behavior For behavioral pattern sub-model parameters, such as user A's P behavior =0.6, P resource To adapt sub-model parameters to resources, such as user A's P resource=0.8; Substituting into the calculation, we get Pglobal=0.5×0.75+0.3×0.6+0.2×0.8=0.715, which is used for subsequent risk assessment.

[0030] By configuring hardware (node ​​server parameters), encryption details (RSA algorithm / parameter format), algorithm selection (sub-model algorithm), and aggregation example (user A parameter calculation), a complete multi-party collaborative processing architecture is implemented. The engineering implementation methods of this architecture, such as node deployment and protocol selection, are clearly defined. At the same time, the privacy protection effect is demonstrated through experimental verification of the Level 3 Information Security Protection Assessment.

[0031] (2) Three-step modeling process: Each step includes formulas, parameters, examples, and performance standards: Similar group data adaptation serves as a cold start solution. The public dataset used is the National Center for Disease Control and Prevention's "National Chronic Disease Surveillance Dataset (2023 Edition)," containing over 100,000 samples covering 31 provinces and municipalities. Features include over 200 dimensions such as age, gender, region, blood pressure, and family medical history. Cosine similarity is used for similarity calculation, selecting only samples with a feature similarity S ≥ 80% with the target user. Sample weights are calculated using the following formula: Among them, W sample Here, is the weight of the common samples, ranging from 0 to 1; e is the natural constant; θ is the weight adjustment coefficient, here θ=2, an empirical value used to control the variation of weights with similarity, determined through 1000 experiments; and S is the feature similarity between the sample and the target user, ranging from 0 to 1. For example, target user A, a 45-year-old male with a family history of hypertension, has a similarity of S=0.92 with the common samples, and W is calculated. sample =1-1 / (1+e 2×0.92The sample with a score of S=0.85 contributes more to the modeling than the sample with S=0.8 and W=0.74, ensuring accurate modeling for new users. Multi-dimensional feature mining is used for key feature extraction, employing the XGBoost algorithm to construct 100 decision trees with a depth of 8 to 12 layers to balance complexity and overfitting. A learning rate of 0.05 to 0.1 ensures convergence and uncovers non-linear correlation features related to physiology, behavior, and genetics. For example, the combined risk of BMI≥28, daily exercise <20 minutes, and a family history of hypertension is 1.8 times that of a single feature. The top 30 core features are selected based on feature importance ranking, i.e., gain value ranking. The cumulative gain value is ≥90%. For example, user A's core features include blood pressure fluctuation amplitude gain=0.25, exercise target achievement rate gain=0.18, and family medical history risk score gain=0.12. Feature calculation examples include blood pressure fluctuation amplitude: the average of the highest and lowest blood pressure per day, where user A's weekly average fluctuation is 19.4 mmHg; and exercise target achievement rate: actual exercise time / recommended exercise time, where user A exercises for 25 minutes daily, achieving a target rate of 83.3%. Feature selection is based on XGBoost's gain value ranking to reduce model computation. Health risk level determination is used as the risk output, employing a formula. Wherein, RR is a health risk value, ranging from 0 to 1. R ≥ 0.7 indicates high risk, 0.3 ≤ R < 0.7 indicates medium risk, and R < 0.3 indicates low risk. The function is the Sigmoid function, where e is the natural constant, mapping the linear combination result to 0 to 1, k is the core feature index (1 to 30), and P... global,k The global weight of the k-th feature, such as P global,1 =0.25, F k For example, the standardized value of the k-th feature, such as user A's blood pressure fluctuation range F1=0.7, b=-0.5 as the bias term, clinically calibrated, risk level is divided into R≥0.7, high risk (≥30% risk of developing chronic disease within 1 year), 0.3≤R<0.7, medium risk (10% to 30%), and R<0.3, low risk (<10%). The effectiveness criterion is to ensure the AUC value is ≥0.88 through five-fold cross-validation; otherwise, the parameters are readjusted. In the example, the standardized values ​​of user A's 30 core features are F1=0.7. k If blood pressure fluctuation F1=0.7 and exercise target achievement rate F2=0.6, and we substitute these values ​​into the formula to get R=0.8, then it is considered high risk.

[0032] By using details of public datasets, weight formulas, feature calculations, and risk examples, the three-step modeling process is transformed into quantifiable and verifiable technical operations. Each step has clear performance standards, such as AUC ≥ 0.88 and cumulative gain ≥ 90%. At the same time, the effectiveness of the cold start solution is verified by achieving a new user accuracy rate of ≥ 75%.

[0033] (3) Initial physical examination cycle generation: The system combines a dual-dimensional cycle generation approach, considering both accessibility to medical resources and risk levels. This includes the source of the basic cycle, resource assessment, cycle calculation, and clinical validation. The basic cycle, based on medical guidelines, serves as the benchmark for claim 5, strictly adhering to the "Expert Consensus on Basic Items of Health Checkups in China (2023 Edition)" and the "Healthy China Action (2023-2030)." Specific standards are 2 years for those aged 18-39, 1 year for those aged 40-59, and 0.5 years for those over 60. For patients with special diseases, such as diabetes and hypertension, adjustments are made according to specific guidelines; for example, 0.5 years for diabetic patients. Basic cycle data is stored in a basic cycle database and adjusted every 3 years according to guideline updates to ensure consistency with clinical standards. The accessibility assessment of medical resources, as a resource dimension, uses a multi-dimensional weighted calculation, with appointment waiting time weighted at 0.4. The longest acceptable waiting time is 30 days, and the weight of equipment usage saturation is 0.3. The calculation assigns a weight of 0.2 to the nearest medical examination institution. The longest acceptable distance is 20 kilometers. For special groups, such as the elderly / disabled, the weight of the transportation convenience dimension is increased by 0.1. If there is direct public transportation, D increases by 0.1 to 0.2. The evaluation formula is M = 0.4T + 0.3U + 0.2D + 0.1C t Where M is the accessibility assessment value for medical resources, ranging from 0 to 1; T is the standardized value for appointment waiting time; U is the standardized value for equipment utilization saturation; D is the standardized value for distance; and C... t For ease of transportation, the range is 0 to 1. In the example, user A is from a rural area and has diabetes. If the waiting time is 7 days, then T=0.77; if the equipment utilization rate is 80%, then U=0.2; if the distance is 5 kilometers, then D=0.75; if there is direct public transportation, then C=0.75. t =1, calculated as M=0.4×0.77+0.3×0.2+0.2×0.75+0.1×1=0.6; the initial cycle calculation and clinical verification are the final output, based on the basic cycle of the "Expert Consensus on Basic Items of Health Check-up in China", such as 1 year for 40 to 59 years old, and 0.5 years for family history of diabetes according to the standard for diabetic patients, combined with the risk level adjustment coefficient K, high risk 0.5 to 0.7, medium risk 1.0, low risk 1.2 to 1.4, and the assessment value M, using formula T init =T base ×K×M, where T init For the initial physical examination cycle, year, T baseThe guideline uses a baseline period, where K is the risk adjustment coefficient. The adjustment coefficient K is categorized by risk level: 0.5 to 0.7 for high risk (R≥0.7), 1.0 for medium risk (0.3≤R<0.7), and 1.2 to 1.4 for low risk (R<0.3). For patients with special diseases, the lower limit of K is increased by 0.1; for example, K=0.6 for high-risk diabetic users. Clinical validation requires meeting the minimum period requirement for special diseases, such as a period of no less than 2 months for diabetic patients. After successful validation, the results are synchronized to the user's mini-program and the hospital's HIS system. In the example, user A is a diabetic patient, and T... base =0.5 years, R=0.8, K=0.6, M=0.6, calculate T init =0.5×0.6×0.6=0.18 years≈2.2 months, which meets the minimum cycle limit for diabetic patients. After synchronization, the hospital will give priority to arranging appointments for them.

[0034] The system comprises basic cycle standards, resource assessment formulas, clinical validation, and example calculations, enabling the complete generation of initial physical examination cycles. It clarifies that the cycle is a dual-dimensional adaptation of risk and resources, rather than being driven by a single risk. Furthermore, clinical validation ensures the rationality of the cycle, avoiding excessive or insufficient physical examinations.

[0035] (4) Quantitative feedback dynamic adjustment mechanism: A mechanism for intervention, feedback, and adjustment has been established, with each step having a protocol library, quantitative formulas, adjustment rules, and coordinated processing, as detailed below: Intervention suggestion delivery serves as the intervention step. The protocol library is categorized by risk level and core characteristics, containing over 100 standardized protocols. For example, the high-risk blood pressure abnormality protocol includes 30 minutes of brisk walking daily (heart rate 100-120 beats / minute), daily salt intake ≤5 grams (the amount measured in a beer bottle cap), blood pressure measurement twice a week (morning / bedtime), and a monthly follow-up. The medium-risk insufficient exercise protocol includes 5 exercise sessions per week, including 3 brisk walks and 2 swimming sessions, each 30 minutes, and quarterly BMI monitoring with a target <24. The protocol library is updated every six months according to the *Chinese Guidelines for the Prevention and Treatment of Hypertension* and the *Chinese Dietary Guidelines* to ensure the clinical effectiveness of the recommendations. Intervention effect evaluation serves as the feedback step, using a three-dimensional weighted calculation, with the degree of improvement of the indicators weighted at 0.4. This only applies to abnormal indicators; the weight of the degree of solution completion is 0.3. The calculation data comes from automatic uploads from wearable devices, with the subjective health score having a weight of 0.3. The calculation is based on a monthly pop-up form in the mini-program, and the evaluation value is calculated using the formula E = 10 × (0.4 × I + 0.3 × C). d +0.3×S), where E is the intervention effect assessment value, ranging from 1 to 10, with E≥8 being excellent, 5≤E<8 being good, and E<5 being poor; I is the degree of improvement of the indicator; C dFor the completion degree of the plan, S is the standardized value of the subjective health score. For abnormal handling, when the device is disconnected, the average value in the past 3 days is used as a temporary substitute. If the user does not fill in the score, the previous score is used as a temporary substitute. After the device is restored, it is recalibrated. In the example, after 3 months of intervention for user A, the blood pressure dropped from 150 / 95 to 135 / 85, and I = 0.75. User A plans to exercise 5 times a week, and the actual completion rate of 4 exercises is 90%, C d = 0.9; The self-evaluation score is 8 points, S = 0.8, and it is calculated that E = 8.1 points, which is excellent; As the adjustment link, the cycle adjustment and linkage optimization use the formula T adj = T init ×(1 + δ), where T adj is the adjusted physical examination cycle, in years, and δ is the adjustment coefficient. For excellent cases, it is 0.1 to 0.2. For patients with special diseases, the range is reduced by 20%. In user A, δ = 0.16. The adjustment coefficient δ is divided according to the effect. For excellent cases, it is 0.1 to 0.2, for good cases, it is 0, and for poor cases, it is -0.15 to -0.3. For patients with special diseases, the range of δ is reduced by 20%. For example, for diabetic patients, the excellent δ = 0.16. The linkage processing includes cycle update, that is, synchronizing to the user and the hospital, and pushing the adjustment notice including the reason, and model optimization, that is, adjusting the data to be sent back to the hybrid model training engine. Every 1000 pieces are accumulated to trigger incremental training. In the example of user A, T init = 2.2 months, E = 10×(0.4×0.75 + 0.3×0.9 + 0.3×0.8) = 8.1 points, where δ = 0.16, and it is calculated that T adj = 2.2×(1 + 0.16) = 2.55 months. After the adjusted data is sent back, the weight of the exercise compliance rate in the model is increased.

[0036] Through the intervention plan library: standardized content, effect evaluation formula: three-dimensional quantification, adjustment linkage: combination of cycle and model, example: user A, 2.55 months, the quantitative feedback dynamic adjustment mechanism is fully realized, and it is clear that this mechanism is the whole process from intervention suggestions, effect evaluation, cycle adjustment to model optimization, rather than a single cycle adjustment. At the same time, through the improvement rate of 82.1% of the indicators of excellent users in the experiment, the effectiveness of the mechanism is verified.

[0037] 3. System architecture: It mainly consists of five major system components, each with clearly defined functions, deployments, interfaces, and methodologies. The cross-domain data fusion platform, as the core component, supports cross-domain data aggregation, enabling the access, alignment, governance, and storage of multi-source data. Deployed on the provincial government cloud (Alibaba Cloud Government Cloud), it meets Level 3 information security standards and is configured with an 8-core CPU, 32GB of RAM, and 10TB of storage, supporting data from 1 million users. Components include an interface adaptation module: HL7FHIR for hospitals, MQTT for wearable devices, and RESTful for government platforms; and a spatiotemporal alignment module. The system includes a time calibration unit that synchronizes UTC time using NTPv4, a spatial mapping unit that converts latitude and longitude using the Gaode API, and a data storage unit that stores data in hot / warm / cold tiers. The data governance module features credibility grading and conflict resolution. The encrypted storage module uses AES256 encryption for sensitive fields and off-site disaster recovery. The time calibration output format of the spatiotemporal alignment module is YYYY-MM-DDHH:MM:SS.fff, and the spatial mapping output format is latitude, longitude, and detailed address. The data storage hot data query latency is ≤10ms, and it supports real-time evaluation of intervention effects.

[0038] The multi-party collaborative processing center, as a key component, supports multi-party collaborative processing, enabling parameter scheduling, encrypted computation, and model aggregation. Deployed on a private cloud within a medical institution to avoid public network risks, it is configured with a 16-core CPU, 64GB of RAM, and 2TB of storage. Its components include a collaborative parameter scheduling module (categorizing and distributing parameters while monitoring transmission progress); an encrypted computation module (RSA decryption of parameters); and a model parameter aggregation module (aggregating parameters, saving versions, and supporting rollback). This center serves as the engineering platform, and its private cloud deployment ensures security. The hybrid model training engine, as the core computing component, supports a three-step modeling process, enabling model training, updating, and... Bias correction is deployed on an AI-dedicated server GPU: NVIDIA A10, configured with a 32-core CPU + 128GB RAM + 4TB storage. The components include a similar group adaptation module, a multi-decision tree feature mining module, a health risk level determination module, and a bias correction module: including a bias detection unit that uses KL divergence to identify biases, an adaptive adjustment unit that handles data / concept drift, and a model version management unit that uses GitLab to save versions and support rollback. The bias correction module calculates KL divergence monthly, triggers an alert if the divergence is >0.2, supplements 1000 recent data points for data drift, and updates the risk threshold for concept drift.

[0039] The cycle optimization and intervention system, as a business component, supports the initial cycle generation and dynamic adjustment, achieving coordination from cycle and intervention to feedback. Deployed in a hybrid cloud environment—a public cloud for user interaction and a private cloud for core logic—it is configured with an 8-core CPU, 16GB of memory, and 1TB of storage. Its components include an initial cycle generation module, an intervention suggestion push module, an effect evaluation and analysis module, an adjustment execution module (including a rule management unit using the Drools engine to solidify conditions and action rules), a data feedback unit that automatically receives cleaned data, and a visualization unit that displays differentiated content to multiple roles. The rules in the adjustment execution module support online updates and verification, data feedback supports anomaly handling, and visualization is differentiated by user / institution / health department. The security and compliance assurance system, as a security component, provides security for all components. Deployed at the front-end nodes of all components, it includes firewalls and WAFs. Its components include a sensitive information encryption module (AES256 encryption for sensitive fields), a role-based access control module (three-level authorization), and an operation audit log module (recording operations and retaining them for ≥2 years), complying with the Personal Information Protection Law and the Data Security Law, and passing the Level 3 Information Security Compliance Assessment to ensure end-to-end security.

[0040] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for dynamically predicting personalized health checkup cycles based on big data, characterized in that, The estimation method includes: An integrated implementation framework is constructed, which includes multi-source data integration, multi-party collaborative processing, three-step modeling and evaluation, and quantitative feedback and adjustment. Each link is connected through encrypted data interaction. Multi-source data integration: Cross-domain health-related data are collected through time calibration and spatial matching. Specifically, this includes structured health records kept by medical institutions, real-time behavioral information collected by wearable devices, family health history information actively reported by users, and information on the accessibility of medical resources released by government platforms. All data is encrypted and stored uniformly. Multi-party collaborative processing stage: A multi-data holder collaborative processing model is adopted, in which each data holder completes data cleaning and feature extraction within its own system and only transmits the encrypted processing results; The three-step modeling and evaluation process involves sequentially implementing three steps: data adaptation for similar groups, multi-dimensional feature mining, and health risk level determination. This process utilizes public health big data to compensate for the lack of new user data and dynamically optimizes the model based on user data. Based on health risk level, the basic physical examination cycle stipulated by authoritative medical guidelines, and the real-time availability of medical resources, we will develop an initial physical examination cycle tailored to each user. Quantitative feedback and adjustment process: Establish a dynamic adjustment mechanism for quantitative feedback, push targeted health intervention suggestions and track the implementation effect, incorporate post-intervention health change information into health risk reassessment, and adjust the physical examination cycle.

2. The method according to claim 1, characterized in that, The process of collecting cross-domain health-related data includes: The real-time behavioral information collected by wearable devices is time-calibrated using internationally unified time as a benchmark to ensure time consistency of data from different sources; then, spatial matching of data is performed by associating the spatial location information of medical resources with administrative division codes and latitude and longitude coordinates. Credibility levels are set according to the data source, with health records provided by medical institutions having the highest credibility and serving as the core benchmark data; the credibility of behavioral information provided by wearable devices is reasonably set according to the device type and is used to supplement behavioral characteristics; the credibility of family health history information declared by users is determined after cross-checking with existing data and is used to improve family health background. When data from different sources are contradictory, the overall result is calculated according to the weight of the credibility of each data point, and the conflicting data is marked for manual verification and confirmation. Secure and encrypted transmission methods are used to transmit various types of data. Wearable device data uses a low-power transmission mode, while important health data uses a real-time transmission mode.

3. The method according to claim 1, characterized in that, Multi-data-holder collaborative processing models include: Each data holder deploys data processing nodes locally to eliminate illogical information by comparing it with medical common sense and identifying abnormal data, and to fill in missing data by referring to similar user data and matching data patterns. After the local data processing node extracts the basic features, it simplifies redundant features and encrypts the feature information, only transmitting the encrypted feature parameters outward. A comprehensive evaluation model consisting of multiple sub-models is constructed. Among them, the medical institution node is responsible for training the sub-model based on physiological indicators, the equipment manufacturer node is responsible for training the sub-model based on behavioral patterns, and the government platform node is responsible for training the sub-model based on resource adaptation. After the training of each sub-model is completed, the model parameters are only summarized and integrated into the global evaluation model, and the original data is not transmitted throughout the process.

4. The method according to claim 1, characterized in that, The three-step modeling process includes: Select a public health dataset that is highly similar to the target user group in terms of demographic characteristics, and adapt a sufficient number of labeled sample features in the dataset to the target user group by adjusting the sample weights. We use a combination of multiple decision trees to mine data features. By setting reasonable decision tree depth and parameter adjustment range, we can mine non-linear correlation features including physiological indicators and lifestyle habits. Then, we sort and filter out the top n core features that have the greatest impact on health risks. The core features are input into the risk assessment model, which outputs a numerical value reflecting the degree of health risk. Based on this value, the risk is divided into three levels: high, medium, and low. The effectiveness of the model is verified by comparing the model's prediction results with the actual health situation. If the effectiveness does not meet the requirements, the model parameters are adjusted until they do.

5. The method according to claim 1, characterized in that, The process of determining the initial physical examination cycle includes: Taking into account factors such as waiting time for physical examination appointments, saturation of physical examination equipment, and distance from the nearest physical examination institution, the appropriate weights are assigned according to the importance of each factor to calculate the assessment results of medical resource accessibility. The basic physical examination cycle is adjusted according to the user's health risk level: the basic physical examination cycle is appropriately shortened for high-risk users, the basic physical examination cycle remains unchanged for medium-risk users, and the basic physical examination cycle is appropriately extended for low-risk users; then, the adjusted cycle is fine-tuned based on the results of the medical resource accessibility assessment to determine the initial physical examination cycle. The established initial physical examination cycle will be simultaneously fed back to the user and the corresponding medical institution.

6. The method according to claim 1, characterized in that, The dynamic adjustment mechanism for quantitative feedback includes: Based on the user's health risk level and core risk characteristics, appropriate health intervention plans are selected from the preset intervention suggestion library and pushed to the user, covering suggestions including exercise, diet, and monitoring; The intervention effect is evaluated by collecting data on users’ implementation of the intervention plan through wearable devices, obtaining subjective health feelings through users’ self-reports, and obtaining follow-up indicators after the intervention through medical institutions. If the intervention is effective, the physical examination cycle should be extended appropriately; if the prognostic effects are good, the physical examination cycle should remain unchanged; if the prognostic effects are poor, the physical examination cycle should be shortened and the intervention plan optimized; the intervention effect and updated user data should be incorporated into the health risk reassessment regularly to dynamically adjust the physical examination cycle.

7. A personalized physical examination cycle dynamic prediction system based on big data, characterized in that, The forecasting system includes: Cross-domain data fusion platform: integrates functional modules including data interface adaptation, spatiotemporal alignment processing, data credibility assessment, conflict resolution and encrypted transmission; it connects to medical institution systems, wearable devices and government platforms through standardized interfaces, uses ID card number plus fingerprint feature encryption value as unique health identifier, and encrypts and stores the collected multi-source data to form a user health information database; Multi-party collaborative processing center: Deployed on a cloud server, it has functions such as collaborative parameter scheduling, encrypted calculation and model parameter aggregation; it receives encrypted feature parameters transmitted by local nodes of each data holder, coordinates the training process of each sub-model, and aggregates and integrates the sub-model parameters into a global evaluation model; Hybrid model training engine: Includes functional modules such as similar group data adaptation, multi-dimensional feature mining, health risk level determination, model performance monitoring and bias correction; loads public health datasets for model pre-training, and dynamically updates the model based on new user data; The cycle optimization and intervention system consists of initial cycle generation, intervention suggestion push, effect evaluation and analysis, and adjustment execution, and is also equipped with a user interaction module; it formulates the initial physical examination cycle based on the global evaluation model, pushes intervention suggestions and collects feedback information, and then adjusts the physical examination cycle and optimizes the model; Security and compliance assurance system: integrates functional modules such as sensitive information encryption, access control and operation log recording; encrypts sensitive health information, sets access permissions according to different roles, and records the entire data operation process.

8. The system according to claim 7, characterized in that, The spatiotemporal alignment processing module of the cross-domain data fusion platform includes: Time calibration unit: Receives time stamps from various data sources and converts them to internationally unified time, correcting data with significant time deviations using appropriate methods; Spatial mapping unit: Converts medical institution addresses and user residential addresses into latitude and longitude coordinates, matches and associates them with the spatial distribution information of medical resources, and uses the matching results for data credibility assessment; Data storage unit: Data is stored using a combination of temporary and long-term storage. Recent real-time behavior information of wearable devices is stored temporarily to ensure query speed, while all historical health information is stored long-term.

9. The system according to claim 7, characterized in that, The bias correction module of the hybrid model training engine includes: Deviation detection unit: Periodically compares the model's prediction results with the actual health status. When the difference between the two exceeds a reasonable range, it determines that the model has a deviation and issues an early warning. Adaptive adjustment unit: If the deviation is caused by changes in data characteristics, the model is optimized by supplementing similar group data or adding new user data; if the deviation is caused by changes in the health risk definition standards, the risk level classification standards and model evaluation rules are updated. Model version management unit: Saves model parameters from different periods and supports model backtracking and recovery; new models must undergo validity verification before being launched, and can only replace old models after verification is passed.

10. The system according to claim 7, characterized in that, The adjustment execution module of the cycle optimization and intervention system includes: Rule Management Unit: Adopts a configurable rule management approach to solidify the correspondence between health risk levels and adjustments to physical examination cycles and intervention recommendations into rules; Data Feedback Unit: Receives intervention execution data uploaded by wearable devices, subjective health scores submitted by users, and follow-up indicator data provided by medical institutions via an interface, and transmits these data to the effect evaluation and analysis module; Visualization unit: Displays users' health risk level, current physical examination cycle, and trend of intervention effect; pushes high-risk user warning information to medical institutions; and provides health management departments with statistical information on the distribution of user health risks and intervention effects within the region.

Images